[steve2005]

The robots (especially the free ones) sometimes do some things that are so bad even someone who learned bridge that day would never do. Recently I found out why a lot of these happen.

The robots of course have a time limit to do anything, with the advanced robot being longer. This is a common practise for any program.

If the robot reached that limit and hasn't calculated what to do it passes if bidding and places smallest legal card if defending or declaring. (my understanding, if this is wrong please correct me)

Yes, you have to have rule for when program runs out of time. But the rule used is sloppy programing. Some programs end if they run out of time, so I suppose these rules are better than stopping.

Bidding rule should be if our side made last bid return partner to lowest spot in our suit including passing. if NT natural then pass ok, if NT something else would be more complicated (so make up some rule passing probably the rule) but if it was Blackwood/KC should already have calculated bid

Rule if declaring or defending and Gib runs out of time. Gib should play what every card was calculated to be best play so far in the calculations.

like duh, playing smallest card!

[diana_eva]

steve2005, on 2017-September-09, 09:58, said:

The robots (especially the free ones) sometimes do some things that are so bad even someone who learned bridge that day would never do. Recently I found out why a lot of these happen.

The robots of course have a time limit to do anything, with the advanced robot being longer. This is a common practise for any program.

If the robot reached that limit and hasn't calculated what to do it passes if bidding and places smallest legal card if defending or declaring. (my understanding, if this is wrong please correct me)

Yes, you have to have rule for when program runs out of time. But the rule used is sloppy programing. Some programs end if they run out of time, so I suppose these rules are better than stopping.

Bidding rule should be if our side made last bid return partner to lowest spot in our suit including passing. if NT natural then pass ok, if NT something else would be more complicated (so make up some rule passing probably the rule) but if it was Blackwood/KC should already have calculated bid

Rule if declaring or defending and Gib runs out of time. Gib should play what every card was calculated to be best play so far in the calculations.

like duh, playing smallest card!

Are you seeing this in the "live" robot tournaments, or in daylongs / challenges? I think there are some bugs with timeouts for daylongs but I never saw one for the live tourneys.

[steve2005]

There are numerous posts of this happening, for a very long time. Have never seen an explanation or correction of this.
I don't have any other example hands.

[jdonn]

This does not get ignored. I don't deal much with play issues, but Uday constantly works on them. Though I don't fully understand the nature of what causes this to happen, I know it's not so simple like it's just one part of the code that you can somehow fix and it's gone. Some of it is also for practical reasons. Meaning, you could theoretically give GIB more thinking time always, but that uses a lot more finite resources, and also would make GIB always play slowly which most people don't want. Also this is very rare - we see it from reports but those are a miniscule percentage of all hands. I have played a ton of hands with GIB over the years, and nothing like this has ever happened to me.

[steve2005]

jdonn, on 2017-September-09, 23:42, said:

This does not get ignored. I don't deal much with play issues, but Uday constantly works on them. Though I don't fully understand the nature of what causes this to happen, I know it's not so simple like it's just one part of the code that you can somehow fix and it's gone. Some of it is also for practical reasons. Meaning, you could theoretically give GIB more thinking time always, but that uses a lot more finite resources, and also would make GIB always play slowly which most people don't want. Also this is very rare - we see it from reports but those are a miniscule percentage of all hands. I have played a ton of hands with GIB over the years, and nothing like this has ever happened to me.

[johnu]

jdonn, on 2017-September-09, 23:42, said:

This does not get ignored. I don't deal much with play issues, but Uday constantly works on them. Though I don't fully understand the nature of what causes this to happen, I know it's not so simple like it's just one part of the code that you can somehow fix and it's gone. Some of it is also for practical reasons. Meaning, you could theoretically give GIB more thinking time always, but that uses a lot more finite resources, and also would make GIB always play slowly which most people don't want. Also this is very rare - we see it from reports but those are a miniscule percentage of all hands. I have played a ton of hands with GIB over the years, and nothing like this has ever happened to me.

Maybe BBO should give GIBs more thinking time if they are paying customers. I would rather wait a few extra seconds and get a reasonable bid/play then get a fast, nonsensical bid/play. But if not for every bid or play, just for those hands where GIB is going to default to some predetermined bid/play like always passing, or playing the lowest card, or discarding random cards.

[gwnn]

I think this is the cause of the 100% wrong defense (again, not a hyperbole, but GIB choosing a card that is mathematically 100% wrong instead of one that is mathematically 100% right - based on no logical inference other than each hand was dealt 13 cards) reported by me in April:
http://www.bridgebas...defense-by-gib/
Uday replied there once so I hope he's still working on it. Stephen Merriman bumped it once but no reply as of yet.

Even a simple confirmation of "GIB usually bases its plays on simulations, but sometimes it runs out of time them and just clicks on random cards with his hair on fire" would be useful.

[diana_eva]

gwnn, on 2017-September-11, 01:33, said:

I think this is the cause of the 100% wrong defense (again, not a hyperbole, but GIB choosing a card that is mathematically 100% wrong instead of one that is mathematically 100% right - based on no logical inference other than each hand was dealt 13 cards) reported by me in April:
http://www.bridgebas...defense-by-gib/
Uday replied there once so I hope he's still working on it. Stephen Merriman bumped it once but no reply as of yet.

Even a simple confirmation of "GIB usually bases its plays on simulations, but sometimes it runs out of time them and just clicks on random cards with his hair on fire" would be useful.

Stephen Tu replicated your bug on his standalone GIB software so I think it can't be a server timeout, it's a real bug in the GIB code and sounds like it's been there forever.

The timeouts I've seen are like steve describes. For example, robot holding AKQJx vs Txxx plays x to x and opps win, or robot passes in the bidding with some monster hand. They also can't be replicated, if you open a teaching table with bots and give them the same action they'll do the right thing. I have not seen those in a few months though, but they were pretty frequent when daylongs and challenges were launched.

[gwnn]

Maybe it's not a server timeout but some other circumstance in which GIB stops simulating.

[smerriman]

jdonn, on 2017-September-09, 23:42, said:

Also this is very rare - we see it from reports but those are a miniscule percentage of all hands.

Could this type of timeout explain the bizarre results I demonstrated in my two posts here? For example, the 7.6% of times it gives the wrong signal when holding a queen, or the 3.3% of times it underleads a doubleton, both with clearcut cases where simulations can't possibly tell it otherwise.

[jdonn]

I am not talking about signal or lead conventions.

[smerriman]

OK yep, understood that, just wondered if there was a potential for it to be the same base cause. I guess it's an unrelated bug in the code. (In fact, since most of the bugs were duplicated at both tables, timeout probably wouldn't make sense anyway. Just interesting how bugs happened such a large proportion of time).

[johnu]

smerriman, on 2017-September-11, 21:41, said:

Just interesting how bugs happened such a large proportion of time

GIB is a buggy and very complex program that needs a major redesign and rewrite. IMO, BBO doesn't have the resources or money to improve GIB up to the level of the best bridge programs currently being actively developed.

[steve2005]

gwnn, on 2017-September-11, 09:25, said:

Maybe it's not a server timeout but some other circumstance in which GIB stops simulating.

It is possible that it is a network issue. A lot of the time this happens it is in a situation Gib really shouldn't time out. Adding time isn't a solution. Computer programs are very fast, if it hasn't figured it out in 2 seconds, 5 won't help.
here is another recent example don't know what tournament it was in. My link
I played in survivor on weekend and had at least 2 of these obvious errors. One someone else was in same suit but had different bidding and did not have the error. So maybe it was an error in the simulation, or from a difference in seed or Gib thinks I had different hand type than other player.

survivor day 1
My link
ducks A trick 1- that is techincally ok
trick 6 j played AQ3 plays 3!
leads T from AT976 ! that is a NT lead. ducks trick 1! (edit) In case not clear Gib underled an ace In a suit contract as opening lead.
My link

[smerriman]

I think those examples are a bit different; for example, in the second case, you're guaranteed to hold the Ace of diamonds or a diamond void based on your bidding, so it doesn't matter what East plays. It's well known that GIB puts far too much emphasis on you having your bid, so any wild bidding causes it to go haywire.

(If you want some easy 100%s, try bidding 3NT in a competitive auction with no stoppers at all and GIB is almost guaranteed to not cash their tricks).

[steve2005]

smerriman, on 2017-September-17, 19:05, said:

I think those examples are a bit different; for example, in the second case, you're guaranteed to hold the Ace of diamonds or a diamond void based on your bidding, so it doesn't matter what East plays.

If I had AD, which it looks like I might. Ducking gives a trick to QD that can never be got. Except perhaps something rare where it gets trick back later.
Covering Q with K forces my supposed Ace and makes Jack good. At best ducking is equal and covering will be better if I have Ax of diamonds.
Maybe a beginner would somehow get it wrong, but I suspect even they would get it right. Nothing in my bidding indicates a void or singleton Ace which are the only cases where ducking is even equal. I signed off short of slam, so defense probably should have an Ace maybe two.

If Gib does not cover on this hand there is something wrong with Gib play algorithm.

The other hand, there is some chance, ducking might work better, but again even a beginner wouldn't duck.

[smerriman]

steve2005, on 2017-September-17, 22:49, said:

Nothing in my bidding indicates a void or singleton Ace which are the only cases where ducking is even equal.

Based on the descriptions you've shown 27+ total points. I tried switching around some cards in the hand editor, and all of the cases I could come up with where you did have the ace resulted in the same number of tricks, even though it looks like the QD is an "extra trick".

Silly yes, but like I said, GIB assumes you have what it thinks you have, no matter how logical that is. That's something that should be looked at, but in a sense unrelated to times when it does something that bidding-based simulations should never tell it to do.

What it should probably do is adjust the simulations to include more cases, but 'weight' them according to how likely it thinks you are to have that hand. Even if it assigns 0.01% to very unlikely hands, that would cause it to break seeming ties better.

[steve2005]

smerriman, on 2017-September-18, 00:44, said:

What it should probably do is adjust the simulations to include more cases, but 'weight' them according to how likely it thinks you are to have that hand. Even if it assigns 0.01% to very unlikely hands, that would cause it to break seeming ties better.

Yes. This was a big problem in early computer programs. Gib is an "early" bridge programs.
So in ties double dummy it doesn.t take best chance to trick opponents. It is either taking random action or action at top or bottom (I think this) of it's tied actions.

[steve2005]

Ducks trick 1 this is in the daylong. so guess another example of bug Diana is talking about.
I don't see how it could time out GIB has only 3 choices at trick 1, Two are equivalent the third can't be better and should be 1 trick worst most of the time.

If Gib is timing out there is a bug in the part of the program handling the simulations.

[steve2005]

another post link is daylong